NexaAI/functiongemma-270m-it-GGUF

FunctionGemma (270M)

Model Description

FunctionGemma is a lightweight, specialized language model from Google built for function calling (tool use).
It is based on the Gemma 3 270M foundation and fine-tuned to reliably translate natural language instructions into structured tool calls using a dedicated formatting/token scheme.

FunctionGemma is designed as a strong base for building agentic systems and is often best used in a tool-calling loop (define tools → generate tool call → execute tool → feed result back).

Features

• Function calling / tool use: produces structured function calls from tool definitions.
• Lightweight (270M): practical for low-latency and edge-friendly deployments.
• More reliable parsing: uses control tokens and formatting patterns for tool definitions/calls/results.
• Ecosystem-friendly: works with common stacks like Hugging Face Transformers.
• Fine-tune ready: intended to be adapted for domain-specific tools and schemas.

Use Cases

• Agent tool orchestration (APIs, database queries, internal actions)
• Mobile/desktop “actions” routing (e.g., calendar, messaging, reminders)
• Structured extraction into JSON arguments for downstream systems
• Deterministic-ish tool routing for assistants (when paired with constraints/validation)
• On-device prototypes where model size matters

Inputs and Outputs

Input:

• Natural language instruction(s)
• Tool definitions (typically JSON schema or structured tool specs)
• Optional conversation history + tool results, following the FunctionGemma formatting

Output:

• A structured tool/function call (tool name + arguments)
• Optionally, follow-up text depending on your prompting/loop design